[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P11] Prediction of Mt. Aso eruption by multi-species large-scale monitoring data analysis

*Minoru Luke Ideno¹, Takeshi Tsuji¹ (1.University of Tokyo)

Keywords:Machine Learning, Volcano, Eruption

The major eruption at Mt Aso on October 8, 2016, caused a large amount of ash fall on the northeast side of Mt. Aso and in surrounding area, as the ejecta was carried by northeasterly winds. Since ash fall caused by eruptions can cause tremendous damage to people living there, advance prediction of eruptions is extremely important from the standpoint of disaster prevention. Currently, however, eruption forecasting is limited to qualitatively determining the degree of danger based on empirical rules, and quantitative forecasting has not been conducted. In this study, we incorporate the knowledge of informatics into the knowledge of volcanology to make quantitative prediction possible using data from seismographs and GPS. Specifically, we used data on seismic wave velocity changes, amplitude of volcanic microtremors, and baseline length changes. As a result, the accuracy of eruption prediction was increased by classifying the data into two categories: features for long-term prediction and features for short-term prediction. The features for long-term prediction were the integral values of GPS and tiltmeter data for detecting volcanic expansion. The short-term forecasting features are based on an anomaly detection approach for features that indicate increased volcanic activity. The combined results showed higher accuracy than when all the features were treated equivalently.